An issue when using "direct_inference" on external private data.

[mileswyn]

Hello! Thanks for your great work SuPreM!

When I tried to infer your pretrained swinUnetR on my private contrast-enhanced CT, I found that the MONAI function "invertd" did not work. The input size of my CT volume is (512, 512, 709). After building the "test_loader" and put the input volume forward through the "val_transform" (which happened in "direct_inference/dataset/dataloader_test"), the image size changed to (248, 207 ,304), since it was resampled with "spacingd" and cropped with "CropForegroundd".
After inferred with "sliding_window_inference", it's prediction should be turned back to size (512, 512, 709) with "invert_transform", which is in line 68 in "inference.py",
BATCH = invert_transform(organ_name,batch,val_transforms).
However, the size of output prediction was still (248, 207 ,304). It seems that the "invertd" does not work in your code.

After debugging, I found it works after changing the output tensor into a MetaTensor.
The original version in "inference.py", line 67, is
batch[organ_name]=pseudo_label_single.cpu()
It should be
batch[organ_name]=MetaTensor(pseudo_label_single.cpu(), meta=image.meta)

Also, "invertd" does not work when ToTensord(keys=["image"]) in val_transform.
So I guess the line 454 in "direct_inference/dataset/dataloader_test.py" should be deleted.

These changes works for me. And I wonder if you could further test it and update your code.
Thanks again.